DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
This Office Action is in response to applicant’s communication filed 30 March 2026, in response to the Office Action mailed 29 December 2025. The applicant’s remarks and any amendments to the claims or specification have been considered, with the results that follow.
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 30 March 2026 has been entered.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-7 and 9-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Carlini et al. ((Certified!!) Adversarial Robustness for Free!, June 2022, pgs. 1-9 – cited in an IDS) in view of Moliner et al. (Realistic Gramophone Noise Synthesis Using a Diffusion Model, June 2022, pgs. 1-8).
As per claim 1, Carlini teaches a computer-implemented method for training a machine-learning network [a classifier can be further trained on the denoised images (pg. 7, “CIFAR-10 ablation”; etc.)], comprising: receiving an input data from a sensor [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (pg. 1, section 1; pg. 4, section 4; etc.); where images are input data from a sensor (camera)], wherein the input data is indicative of image information, radar information, sonar information, or sound information [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (pg. 1, section 1; pg. 4, section 4; etc.); which are indicative of image information]; generating a training data set utilizing the input data, wherein the generating includes creating one or more copies of the input data and adding noise with a same mean and a known variance [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.); where the fixed mean and variances are “a same mean” and “a known variance”]; utilizing a diffusion model, reconstructing and purifying the training data set by removing noise associated with the input data and reconstructing the one or more copies of the training data set to create a modified input data set [the diffusion denoiser model is applied to the noise-added (training) data xt* to obtain denoised image samples (modified input data set) x̂ (pg. 3, Algorithm 1: step 3; section 3: “Denoised smoothing via a diffusion model”, equation (7); etc.)], wherein removing the noise includes denoising noise, via the diffusion model, back through time the one or more copies across a plurality of timesteps [the diffusion denoiser model is applied to the noise-added (training) data xt* to obtain denoised image samples (modified input data set) x̂ over a number of timesteps (pg. 3, Algorithm 1: steps 1-3; section 3: “Denoised smoothing via a diffusion model”, equation (7); etc.); which can also be applied to whole datasets (pg. 4, section 4; etc.)]; and utilizing a fixed classifier, output a classification associated with the input data in response to a majority vote of the classification obtained by the fixed classifier of the modified input data set [the denoised image is classified with an off-the-shelf (fixed) classifier to produce classification y (pg. 3, Algorithm 1: steps 4-5; section 3: “Denoised smoothing via a diffusion model”, equation (8); etc.); Examiner’s Note: when one classification is received from one classifier, that is a majority vote of the classification (1/1); but Carlini also teaches that a majority vote of the labels predicted by the base classifier can be taken, and if the correct class is output sufficiently often, then the defense’s output on the original un-noised input is guaranteed to be robust to ℓ2 norm bounded adversarial perturbations (pg. 1, section 1; pg. 2, section 2: “Randomized Smoothing”; pg. 4, section 4: “we perform randomized smoothing”; etc.)].
While Carlini teaches using a fixed or learned variances, as well as specified perturbations (see above as well as, e.g., pg. 2, section 2), it has not been relied upon for teaching wherein each of the copies are generated utilizing a noise variance schedule that includes an associated perturbation and severity level for each of the copies.
Moliner teaches wherein each of the copies are generated utilizing a noise variance schedule that includes an associated perturbation and severity level for each of the copies [the forward diffusion model uses a learned noise variance parameter with a predefined schedule (pg. 2, section 3.1; etc.) and the reverse diffusion process includes a noise variance schedule with the samples (pg. 3, section 3.2; etc.), which can also include a perturbation guide (pg. 4, section 4.2) and tail parameters indicating severity of the noise (severity level) (pg. 2, section 2.3; etc.); to generate samples (pg. 1, abstract; pg. 2, sections 3-3.1; etc.)].
Carlini and Moliner are analogous art, as they are within the same field of endeavor, namely using diffusion models including denoising and sample synthesis.
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to utilize the noise variance schedule with each copy, as taught by Moliner, for the noise variance parameters of copies in the system taught by Carlini.
Moliner provides motivation as [the associated noise variance schedule can be modified without needing to retrain the model, providing added flexibility (pg. 3, section 3.2; etc.)].
As per claim 2, Carlini/Moliner teaches wherein the diffusion model and the fixed classifier are both pre-trained [we can use pretrained diffusion models and image classifiers (Carlini: pg. 1, abstract; etc.)].
As per claim 3, Carlini/Moliner teaches wherein the method includes, for each training data set, computing a clean image utilizing the diffusion model and the fixed classifier [the diffusion denoiser model is applied to the noise-added (training) data xt* to obtain a denoised image samples x̂ (Carlini: pg. 3, Algorithm 1: step 3; section 3: “Denoised smoothing via a diffusion model”, equation (7); etc.) and the denoised image is classified with an off-the-shelf (fixed) classifier to produce classification y (Carlini: pg. 3, Algorithm 1: steps 4-5; section 3: “Denoised smoothing via a diffusion model”, equation (8); etc.); where the denoiser produces a (as close as possible) clean image (Carlini: pg. 2, section 2: “Denoising Diffusion Probabilistic Models”; etc.)].
As per claim 4, Carlini/Moliner teaches wherein the noise includes Gaussian noise, shot noise, motion blur, zoom blur, compression, or brightness changes [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.)].
As per claim 5, Carlini/Moliner teaches wherein the fixed classifier and diffusion model are trained on a same data distribution [the diffusion denoise model and pretrained classifier can be finetuned using the same training set (Carlini: pgs. 5-7, section 4.1; etc.)].
As per claim 6, Carlini/Moliner teaches wherein the diffusion model is configured to reverse noise associated with the training data set by denoising through time [the diffusion model can be applied iteratively to reverse noise through time (Carlini: pgs. 3-4, section 3, “Denoised smoothing via a diffusion model”)].
As per claim 7, Carlini/Moliner teaches wherein the diffusion model is denoised [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.)].
As per claim 9, Carlini teaches a system including a machine-learning network, comprising: an input interface configured to receive input data from a sensor, wherein the sensor includes a camera, a radar, a sonar, or a microphone [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (pg. 1, section 1; pg. 4, section 4; etc.); where images are input data from a sensor (camera)]; and a processor in communication with the input interface [classification is performed on an A100 GPU (pg. 5, section 4; etc.); which includes a processor (GPU) in communication with an input interface], wherein the processor is programmed to: receive the input data from the input interface [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (pg. 1, section 1; pg. 4, section 4; etc.) where the classification is performed on an A100 GPU (pg. 5, section 4; etc.); which includes a processor (GPU) in communication with an input interface], wherein the input data is indicative of image, radar, sonar, or sound information [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (pg. 1, section 1; pg. 4, section 4; etc.); which are indicative of images]; generate a training data set utilizing the input data, wherein the training data set includes with a number of copies of the input data along with noise [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.)]; reconstruct and purify the training data set by removing the noise associated with the input data and reconstructing the number of copies to create a modified input data set [the diffusion denoiser model is applied to the noise-added (training) data xt* to obtain denoised image samples (modified input data set) x̂ (pg. 3, Algorithm 1: step 3; section 3: “Denoised smoothing via a diffusion model”, equation (7); etc.)] wherein removing the noise includes denoising the noise, via the diffusion model, back through time the one or more copies across a plurality of timesteps [the diffusion denoiser model is applied to the noise-added (training) data xt* to obtain denoised image samples (modified input data set) x̂ (pg. 3, Algorithm 1: step 3; section 3: “Denoised smoothing via a diffusion model”, equation (7); etc.)]; and output a final classification associated with the input data in response to a majority vote of classifications obtained from the modified input data set [the denoised image is classified with an off-the-shelf (fixed) classifier to produce classification y (pg. 3, Algorithm 1: steps 4-5; section 3: “Denoised smoothing via a diffusion model”, equation (8); etc.); Examiner’s Note: when one classification is received from one classifier, that is a majority vote of the classification (1/1); but Carlini also teaches that a majority vote of the labels predicted by the base classifier can be taken, and if the correct class is output sufficiently often, then the defense’s output on the original un-noised input is guaranteed to be robust to ℓ2 norm bounded adversarial perturbations (pg. 1, section 1; pg. 2, section 2: “Randomized Smoothing”; pg. 4, section 4: “we perform randomized smoothing”; etc.)].
While Carlini teaches using a fixed or learned variances, as well as specified perturbations (see above as well as, e.g., pg. 2, section 2), it has not been relied upon for teaching wherein each of the copies are generated utilizing a noise variance schedule that includes an associated perturbation and severity level for each of the copies.
Moliner teaches wherein each of the copies are generated utilizing a noise variance schedule that includes an associated perturbation and severity level for each of the copies [the forward diffusion model uses a learned noise variance parameter with a predefined schedule (pg. 2, section 3.1; etc.) and the reverse diffusion process includes a noise variance schedule with the samples (pg. 3, section 3.2; etc.), which can also include a perturbation guide (pg. 4, section 4.2) and tail parameters indicating severity of the noise (severity level) (pg. 2, section 2.3; etc.); to generate samples (pg. 1, abstract; pg. 2, sections 3-3.1; etc.)].
Carlini and Moliner are analogous art, as they are within the same field of endeavor, namely using diffusion models including denoising and sample synthesis.
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to utilize the noise variance schedule with each copy, as taught by Moliner, for the noise variance parameters of copies in the system taught by Carlini.
Moliner provides motivation as [the associated noise variance schedule can be modified without needing to retrain the model, providing added flexibility (pg. 3, section 3.2; etc.)].
As per claim 10, see the rejection of claim 4, above.
As per claim 11, Carlini/Moliner teaches wherein the input data is indicative of an image [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (Carlini: pg. 1, section 1; pg. 4, section 4; etc.)], and the training data set is generated by selecting each pixel associated with the image randomly drawn from a Gaussian distribution [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and “The diffusion process transforms images from the target data distribution to purely random noise over time. The reverse process then synthesizes images from the data distribution starting with random Gaussian noise.” (Carlini: pg. 2, “Denoising Diffusion Probabilistic Models”); where adding random Gaussian noise is selecting pixels of the image randomly from a Gaussian distribution].
As per claim 12, Carlini/Moliner teaches wherein the system includes a diffusion model that is a denoised diffusion model configured to generate images through a diffusion process [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.)].
As per claim 13, Carlini/Moliner teaches wherein the diffusion model is utilized to reconstruct and purify the training data set [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.); which is reconstructing and purifying the (noise-added) training data set].
As per claim 14, Carlini/Moliner teaches wherein the final classification is output utilizing a classifier [the denoised image is classified with an off-the-shelf (fixed) classifier to produce classification y (Carlini: pg. 3, Algorithm 1: steps 4-5; section 3: “Denoised smoothing via a diffusion model”, equation (8); etc.)].
As per claim 15, see the rejection of claim 1, above, wherein Carlini/Moliner also teaches a computer-program product storing instructions which, when executed by a computer, cause the computer to: [perform the method] [the method can be implemented in code (Carlini: pg. 3, Algorithm 1; etc.) stored in memory and executed by a GPU (computer) (Carlini: pg. 5, section 4; etc.)].
As per claim 16, Carlini/Moliner teaches wherein the input data includes an image, radar, sonar, or sound information [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (pg. 1, section 1; pg. 4, section 4; etc.)].
As per claim 17, Carlini/Moliner teaches wherein adding noise includes adding noise with a same mean and a same variance to each of the one or more copies [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.); where the fixed mean and variances are “a same mean” and “a same variance”].
As per claim 18, Carlini/Moliner teaches wherein adding noise includes adding noise with a same mean [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.); where the fixed mean and variances are “a same mean” and “a same variance”].
As per claim 19, Carlini/Moliner teaches wherein adding noise includes adding noise with a same variance [“A forward diffusion process takes a source data distribution (e.g., images from some data distribution) and then adds Gaussian noise until the distribution converges to a high-variance isotropic Gaussian. Denoising diffusion models are trained to invert this process. Thus by intervening in the reverse diffusion process, we can use a diffusion model as a denoiser that recovers high quality denoised inputs from inputs perturbed with Gaussian noise” (Carlini: pg. 1, section 1; etc.) and noise is added to the input images with fixed mean and fixed (or learned) variances βt, where xt* shows the (training) data set utilizing a clean training image x (input data) with added noise (Carlini: pgs. 2-3, section 2: “Denoising Diffusion Probabilistic Models”; Algorithm 1: step 2; and section 3: “Denoised smoothing via a diffusion model” and equations (4) and (6); etc.); where the fixed mean and variances are “a same mean” and “a same variance”].
Claim(s) 8 and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Carlini et al. ((Certified!!) Adversarial Robustness for Free!, June 2022, pgs. 1-9 – cited in an IDS), in view of Moliner et al. (Realistic Gramophone Noise Synthesis Using a Diffusion Model, June 2022, pgs. 1-8), and further in view of Wexler (WO 2022/130011).
As per claim 8, Carlini/Moliner teaches wherein the sensor is a camera [the system uses ImageNet (or CIFAR-10) images as its input dataset for classification (Carlini: pg. 1, section 1; pg. 4, section 4; etc.); where images are input data from a sensor (camera)].
While Carlini/Moliner teaches that the input data includes images (see above) it has not been relied upon for teaching the input data includes video information obtained from the camera.
Wexler teaches the input data includes video information obtained from the camera [the system receives one or more image signals output from a camera and provides a classification (paras. 0006-7; see also: 0870, 0887-888, etc.) which image data can include video clips from a video camera (paras. 0259, 0295; see also: 0152, 0160, 0164, 0312, 0578, etc.)].
Carlini/Moliner and Wexler are analogous art, as they are within the same field of endeavor, namely classifying images using a classifier.
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to include video camera images in the images being classified by the classifier system, as taught by Wexler, in the images being classified by the classifier system taught by Carlini/Moliner.
Because both Carlini/Moliner and Wexler teach image classification systems, including classifying noisy data, it would have been obvious to one of ordinary skill in the art to include video camera images in the images being classified by the classifier system, as taught by Wexler, in the images being classified by the classifier system taught by Carlini/Moliner, to achieve the predictable result of expanding the range of inputs on which the classifier can be used/trained.
As per claim 20, Carlini/Moliner teaches the computer-program product of claim 15, as described above.
Carlini/Moliner has not been relied upon for teaching wherein the input data includes sound information obtained from a microphone.
Wexler teaches wherein the input data includes sound information obtained from a microphone [the classifier can provide context classification to both image inputs and sound inputs taken from a microphone (paras. 0870, 0887-888, etc.)].
Carlini/Moliner and Wexler are analogous art, as they are within the same field of endeavor, namely using a classifier, including classifying images and noisy data.
It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to include video camera images and microphone sound data in the input data being classified by the classifier system, as taught by Wexler, in the inputs being classified by the classifier system taught by Carlini/Moliner.
Because both Carlini/Moliner and Wexler teach image classification systems, including classifying noisy data, it would have been obvious to one of ordinary skill in the art to include video camera images and microphone sound data in the input data being classified by the classifier system, as taught by Wexler, in the inputs being classified by the classifier system taught by Carlini/Moliner, to achieve the predictable result of expanding the range of inputs on which the classifier can be used/trained, as well as allowing further context for classification (see, e.g., Wexler: paras. 0870, 0887-888, etc.).
Response to Arguments
The prior claim objections have been withdrawn due to the amendments filed.
The prior rejections under 35 U.S.C. 112 have been withdrawn due to the amendments filed.
Applicant's arguments filed 30 March 2026 have been fully considered but they are not persuasive.
Applicant argues that the cited art only teaches denoising one sample at a time, “rather than performing dataset-level reconstruction and transformation.”
However, Carlini teaches the diffusion denoiser model is applied to the noise-added (training) data xt* to obtain denoised image samples (modified input data set) x̂ over a number of timesteps (pg. 3, Algorithm 1: steps 1-3; section 3: “Denoised smoothing via a diffusion model”, equation (7); etc.); which can also be applied to whole datasets (pg. 4, section 4; etc.). Additionally, even if the system/method taught by Carlini were only used to denoise individual samples of the training dataset, this would still be within the broadest reasonable interpretation of the claimed “reconstructing and purifying the training dataset by removing noise…”
Conclusion
The following is a summary of the treatment and status of all claims in the application as recommended by M.P.E.P. 707.07(i): claims 1-20 are rejected.
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Cohen et al. (Certified Adversarial Robustness via Randomized Smoothing, June 2019, pgs. 1-36 – also cited by Carlini) – discloses adding Gaussian noise to input images, inputting the noisy images into a classifier, and taking a majority vote for classification (“Randomized Smoothing”).
Salman et al. (Denoised Smoothing: A Provable Defense for Pretrained Classifiers, Sept 2020, pgs. 1-29 – also cited by Carlini) – discloses a system similar to Carlini, above, but also including training the denoiser before applying denoised samples to the pretrained classifier.
Nie et al. (Diffusion Models for Adversarial Purification, May 2022, pgs. 1-22 – cited in an IDS, and also cited by Carlini) – discloses taking an adversarial example (noise-added), and applying a diffusion model for purifying the example for classification, similar to Carlini, above.
Ho et al. (Denoising Diffusion Probabilistic Models, Dec 2020, pgs. 1-25 – cited in an IDS, and also cited by Carlini) – discloses a denoising diffusion probabilistic model, such as that used above, used for image synthesis.
Kandpal (US 2023/0343312) – discloses applying a denoising diffusion probabilistic model to enhanced/noisy training data, as a vocoding module on sound data.
Huang (US 2023/0153949) – discloses an image generation system including training of a denoising diffusion probabilistic model as the generative model, and utilizing training data with Gaussian noise added.
Saharia (US 2023/0067841) – discloses an image generation system including training of a denoising diffusion probabilistic model as the generative model, utilizing training data with Gaussian noise added, as well as using the denoised output of the diffusion model for improving downstream image classification model(s) and specifying a noise variance of forward process steps (a schedule).
Lecuyer et al. (Certified Robustness to Adversarial Examples with Differential Privacy, May 2019, pgs. 1-18) – discloses training the model including alternating between SGD steps with fixed noise variance and projection steps.
The examiner requests, in response to this Office action, that support be shown for language added to any original claims on amendment and any new claims. That is, indicate support for newly added claim language by specifically pointing to page(s) and line number(s) in the specification and/or drawing figure(s). This will assist the examiner in prosecuting the application.
When responding to this office action, Applicant is advised to clearly point out the patentable novelty which he or she thinks the claims present, in view of the state of the art disclosed by the references cited or the objections made. He or she must also show how the amendments avoid such references or objections. See 37 CFR 1.111(c).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GEORGE GIROUX whose telephone number is (571)272-9769. The examiner can normally be reached M-F 10am-6pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Omar Fernandez Rivas can be reached at 571-272-2589. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/GEORGE GIROUX/Primary Examiner, Art Unit 2128
Prosecution Insights